Nat Lambert reposts his lecture on over-optimization and reward hacking
natolambert · x · 2026-07-25
A repost of Nat Lambert’s shorter lecture on over-optimization, reward hacking, sycophancy, and verbosity.
The lecture revisits Goodhart’s Law, discusses why rubrics can be over-optimized in the same way as reward models, and frames RLVR as a separate phenomenon. It also touches on misalignment signals, style-versus-substance critiques, and the Llama 4 leaderboard-gaming example.
Related event: Lectures Explore LLM Over-Optimization and Reward Hacking(3 posts)→
More from AGI Musings
- ARC AGI 3 should have stayed private, with no examples or public dataset — flowersslop · 2026-07-27
- AI makes knowledge cheaper, but judgment remains the scarce skill — _jaydeepkarale · 2026-07-27
- A reply to François Chollet argues intelligence is only part of the advantage — binarybits · 2026-07-27
- A Chollet quote reignites the debate over whether intelligence has diminishing returns — binarybits · 2026-07-27
- Noahpinion quotes Chollet: intelligence may hit a hard ceiling — binarybits · 2026-07-27
- Using LLMs to optimize the next LLM is not singularity, argues Burkov — burkov · 2026-07-27